Heating, ventilating and air conditioning energy-saving optimization system for indoor ski field

Abstract

The embodiment of the invention provides an energy-saving optimization system for heating, ventilating and air conditioning of an indoor ski field. The energy-saving optimization system comprises a multi-source sensing layer, an edge computing layer, a cloud decision-making layer and an equipment execution layer. The multi-source sensing layer is used for collecting multi-source data such as weather, passenger flow, temperature and humidity and equipment state; the edge calculation layer carries out fusion processing on the data and generates a load prediction result through a load prediction mechanism; the cloud decision-making layer generates an optimization control instruction based on a multi-agent deep reinforcement learning and model prediction control optimization strategy; and the equipment execution layer receives and executes the instruction and feeds back the equipment state. Through a multi-layer collaborative optimization architecture, accurate load prediction and equipment intelligent collaborative control are realized, five-stage stepped optimization and a dynamic priority mechanism are adopted, the energy efficiency of the system is remarkably improved, the energy consumption is reduced while the environmental comfort is ensured, and the economical efficiency and the stability of system operation are effectively improved.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the field of heating, ventilation and air conditioning energy-saving control, and particularly relates to an indoor ski field heating, ventilation and air conditioning energy-saving optimization system. BACKGROUND

[0002] With the rise of ice and snow sports worldwide, indoor ski fields as a place for ice and snow sports that can operate all year round have been widely concerned and developed. However, due to the special building structure and functional requirements of the indoor ski field, the energy consumption of the heating, ventilation and air conditioning system is extremely high. The system not only needs to maintain a low-temperature environment in the snow area to meet the skiing requirements, but also needs to ensure that the audience area, service area and other non-skiing areas are in a comfortable temperature range. Therefore, how to realize the energy-saving optimization of the heating, ventilation and air conditioning system of the indoor ski field, reduce its operating cost, and at the same time ensure the comfort of the indoor environment, has become an important problem to be solved in the current operation of the indoor ski field.

[0003] In the prior art, the heating control of the indoor ski field mainly adopts the following implementation ways: in terms of load prediction, the basic prediction method relies on historical energy consumption data or a single environmental parameter to construct a prediction model, and some schemes introduce simple machine learning models such as time series and neural networks, but do not take into account dynamic factors such as outdoor weather and passenger flow fluctuations, and do not establish a multi-source data fusion mechanism and a weight dynamic adjustment strategy; in terms of device control, a model predictive control (MPC) scheme is used to construct a static system model for optimization, or a simple partition and independent control strategy is used to uniformly adjust the parameters of or independently optimize a single device such as a cold machine, a water pump and a fan, or it relies on artificial experience control to achieve by presetting a fixed operation mode and manually adjusting; in terms of regional optimization and execution strategy, the ski field is usually simply divided into two independent control units of a skiing area and a supporting area, adjustment is triggered based on real-time temperature feedback, and a direct joint optimization mode of the whole system is adopted, without setting a hierarchical optimization strategy and a dynamic priority mechanism.

[0004] The prior art mainly has the following problems: the load prediction accuracy is low and the self-adaptive ability is insufficient, key dynamic data such as weather and passenger flow are not fused, resulting in that the prediction lags behind the actual load change; device coordination is missing and global optimization is weak, device control adopts a "one-size-fits-all" parameter adjustment, MPC and multi-agent are not effectively combined, and local optimization leads to an increase in global energy consumption; regional subdivision is insufficient and the execution strategy is lagging, the internal thermal environment difference of the region is not considered, the control logic relies on real-time feedback rather than prediction data, the load fluctuation cannot be perceived in advance, and problems such as frequent system start-stop, high energy consumption, poor control accuracy and accelerated device wear are caused. These defects eventually lead to serious consequences such as high energy consumption, insufficient environmental control accuracy and fast device wear of the existing system. SUMMARY

[0005] The embodiment of the present application provides an energy-saving optimization system for a heating, ventilation and air conditioning (HVAC) system of an indoor ski field to solve the above problems in the prior art.

[0006] The embodiment of the present application provides an energy-saving optimization system for a heating, ventilation and air conditioning (HVAC) system of an indoor ski field, comprising: A multi-source perception layer is configured to collect multi-source data of an indoor ski field environment, wherein the multi-source data comprises meteorological data, passenger flow data, temperature and humidity data and equipment state data; An edge computing layer is configured to perform data fusion processing on the multi-source data and generate a load prediction result through a load prediction mechanism; A cloud decision layer is connected to the edge computing layer and configured to receive the load prediction result transmitted by the edge computing layer and the equipment state fed back by an equipment execution layer, generate an optimized control instruction based on a multi-agent deep reinforcement learning and a model predictive control optimization strategy; The equipment execution layer is connected to the cloud decision layer and comprises HVAC execution equipment, configured to receive and execute the optimized control instruction and feed back the equipment state to the edge computing layer and the cloud decision layer.

[0007] Preferably, the multi-source perception layer comprises a meteorological data collection unit, a passenger flow data collection unit, a temperature and humidity data collection unit and an equipment state data collection unit; The meteorological data collection unit is configured to collect outdoor temperature, humidity and solar radiation intensity data; The passenger flow data collection unit is configured to collect passenger flow data through gate counting, Wi-Fi probe MAC address identification and video monitoring image analysis; The temperature and humidity data collection unit is distributed in different areas of the indoor ski field and configured to collect temperature and humidity data of each area; The equipment state data collection unit is configured to collect operating parameters of a refrigeration unit, a cooling tower, a circulating water pump and a damper.

[0008] Preferably, the edge computing layer comprises a load prediction module, and the load prediction module comprises a data preprocessing unit, a hybrid deep learning model construction unit and a dynamic weight self-adaptive unit; wherein, The data preprocessing unit is configured to receive the multi-source data collected by the multi-source perception layer, detect and eliminate outliers by using an isolation forest algorithm, remove high-frequency noise by using a moving average filter, and supplement missing data by using a linear interpolation method; The hybrid deep learning model construction unit adopts a long short-term memory network or a Transformer model as a core framework, and the input features include climate features, passenger flow features and historical load features, wherein the climate features are obtained by processing meteorological data, the passenger flow features are obtained by processing passenger flow data, and the historical load features are obtained by extracting historical energy consumption data; The dynamic weight adaptive unit adopts a lightweight convolutional neural network to dynamically adjust the historical data weight, real-time meteorological data weight and real-time passenger flow data weight according to real-time disturbance characteristics, and obtain a load prediction result through weighted fusion of the three.

[0009] As preferred, the mixed deep learning model construction unit selects historical energy consumption data and corresponding meteorological, passenger flow and temperature and humidity data of a time period to form a training data set in the training stage, divides the training data set into a training set and a validation set, iteratively optimizes model parameters using an Adam optimizer, takes root mean square error as a loss function of model training, and converges the model loss value through multiple iterations. The mixed deep learning model construction unit adjusts the model training focus in combination with working condition data corresponding to different seasons and different passenger flow intensities in the training process, and ensures stable output of the load prediction result under various working conditions.

[0010] As preferred, the real-time disturbance characteristics include a passenger flow anomaly index and a weather change rate. The passenger flow anomaly index is obtained by calculating the deviation rate of the current passenger flow data from the historical same-period passenger flow mean value. The weather change rate is obtained by calculating the change slope of the outdoor temperature in a set period. The dynamic weight adaptive unit adjusts the proportion of the historical data weight, real-time meteorological data weight and real-time passenger flow data weight according to the change of the real-time disturbance characteristics, so that the load prediction result after weighted fusion adapts to the current working condition.

[0011] As preferred, the cloud decision layer includes a collaborative optimization control module, and the collaborative optimization control module includes a multi-agent deep reinforcement learning unit and a model predictive control optimizer. The multi-agent deep reinforcement learning unit models the chillers, cooling towers and circulating water pumps at the end of the wind valve as independent agents, the state space of each agent includes its own operating parameters and environmental parameters of the corresponding area, the action space is the adjustable control parameter range of itself, the reward function integrates the energy consumption cost and comfort-related parameters, and the agents share environmental interaction data through an experience replay pool. The model predictive control optimizer adopts a linear time-varying (LTV) model, the state variables of the linear time-varying (LTV) model include the total cooling capacity of the system, the temperature of each region and the energy consumption of the equipment, the input variable is the load prediction result output by the edge computing layer, the output variable is the target control parameter corresponding to each agent, and the key parameters of the linear time-varying model are obtained by combining offline identification and online updating, the initial parameters are identified based on historical operation data in the offline stage, and the model is dynamically updated by receiving the parameter correction value fed back by the multi-agent deep reinforcement learning unit in the online stage.

[0012] As preferred, the synergistic optimization control module further comprises a synergistic working unit; the synergistic working unit is used to realize the synergistic work of the multi-agent deep reinforcement learning unit and the model predictive control optimizer; The load curve output by the load prediction module is used as the feedforward disturbance input of the model predictive control optimizer; the multi-agent deep reinforcement learning unit perceives the system state in real time, learns the device characteristic changes on line through the Actor-Critic mechanism, and outputs the model parameter correction value to the model predictive control optimizer; The model predictive control optimizer solves the optimal control sequence based on the updated linear time-varying LTV model, and issues the control sequence to each agent as a reference constraint; Each agent combines its own learning experience to fine-tune the local strategy within the reference constraint range, generates an execution-level control instruction, and simultaneously feeds back the actual running data to the experience replay pool for the next round of learning optimization.

[0013] As preferred, the edge computing layer further comprises a ladder optimization execution module, and the ladder optimization execution module comprises a five-stage optimization strategy; the five-stage optimization strategy comprises: A single-device independent optimization stage, based on the energy efficiency curve of the refrigeration unit and the real-time load demand, the optimal load rate is found to adjust the output power, and based on the similarity law, the circulating water pump frequency is adjusted according to the actual flow demand to make it run in the high-efficiency zone; A small system synergistic optimization stage, the cooling tower system of the refrigeration host water pump is optimized through a flow temperature difference matching algorithm to optimize the synergistic relationship, and the cooling tower fan frequency and water pump frequency are dynamically adjusted to reduce the total energy consumption of the small system; A regional end optimization stage, based on real-time temperature and humidity sensor data, the end air supply valve opening degree and reheater power are adjusted through a proportional-integral-derivative PID control algorithm to ensure the uniformity of temperature and humidity in the region; A whole system global optimization stage, all subsystem data are integrated, and based on the global optimal target output by the multi-agent deep reinforcement learning and model predictive control optimization module, the linear programming is used to balance the regional load distribution to realize the minimization of the total energy consumption of the system; An economic fine-tuning decision stage, real-time electricity price signals and device life loss models are introduced, and the control strategy is fine-tuned through cost-benefit analysis on the basis of global optimization.

[0014] As preferred, the ladder optimization execution module further comprises a dynamic priority mechanism execution unit; the dynamic priority mechanism execution unit is used to dynamically adjust the optimization strategy according to the real-time monitored passenger flow state and time period characteristics: When the passenger flow density of the ski area is greater than a set threshold, a high load response mode is triggered, the whole system global optimization and the economic fine-tuning decision stage are preferentially executed, and the control cycle is shortened; During the night low load period, switch to the energy saving mode, focus on single device independent optimization and small system collaborative optimization stage, prolong the control cycle.

[0015] Preferably, the step optimization execution module further comprises a closed-loop rolling optimization unit; the closed-loop rolling optimization unit is used for: Through the Internet of Things system, the optimized control instruction is issued to the field executor, and the actual action state fed back by the executor is received; Periodically compare the deviation of the actual running data and the optimization target, trigger the emergency correction mechanism when the deviation exceeds the set threshold, re-execute the optimization process of the corresponding stage, and realize closed-loop control.

[0016] Compared with the prior art, the indoor skiing field heating and air conditioning energy saving optimization system provided by the embodiment of the present application has the following beneficial effects: (1) The present application fuses meteorological, passenger flow, equipment state and other multi-source data, combines a hybrid deep learning model and a dynamic weight self-adaptive mechanism in the edge computing layer, and the system can dynamically fuse historical data and real-time disturbance characteristics, accurately predict future load change trend, effectively solve the prediction lag problem caused by single data and static model in traditional method, and provide accurate data basis for system optimization and regulation.

[0017] (2) The present application adopts the optimization strategy of combining multi-agent deep reinforcement learning and model predictive control in the cloud decision layer, constructs key devices such as refrigeration units, water pumps and cooling towers into intelligent agents with autonomous decision and cooperation ability, realizes dynamic cooperation and global energy consumption optimization among devices, and solves the problems of energy consumption conflict and low overall energy efficiency caused by independent operation of devices in traditional control.

[0018] (3) The five-stage optimization strategy and dynamic priority mechanism realized by the step optimization execution module of the present application, the system can adaptively adjust the optimization emphasis and control cycle according to the working condition changes such as real-time passenger flow density and time period characteristics, realize hierarchical fine regulation from device level to system level, and improve the response speed, stability and economy of the system under different running scenes. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiment or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0020] Figure 1A system architecture diagram of the energy-saving optimization system of the indoor skiing field heating ventilation air conditioning provided by the embodiment of the present application is shown in FIG. 1. Figure 2 A schematic diagram of the MADRL-MPC collaborative mechanism provided by the embodiment of the present application is shown in FIG. 2. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0022] The existing indoor skiing field heating ventilation air conditioning control technology mainly relies on load prediction methods based on historical energy consumption or a single environmental parameter, and device regulation and control strategies such as model predictive control or simple partition independent control. These technologies are difficult to effectively integrate multi-dimensional dynamic data such as outdoor weather changes and passenger flow fluctuations, and have insufficient real-time sensing and prediction capabilities for system load. On the one hand, load changes are affected by multiple factors such as weather, passenger flow and device state, and the existing technology cannot accurately predict and respond to systemic problems such as energy consumption fluctuations and comfort level decline caused thereby; on the other hand, there is a lack of optimization mechanism for device collaboration and overall energy efficiency, and the control strategy cannot be adjusted adaptively according to real-time working conditions, so the optimization chain from sensing to decision-making to execution is broken, making it difficult for the heating ventilation system to achieve efficient energy saving and stable control in the complex and variable indoor skiing field environment.

[0023] The present application aims to provide an energy-saving optimization system of an indoor skiing field heating ventilation air conditioning, which collects multi-source data through a multi-source sensing layer, performs data fusion and load prediction through an edge computing layer, generates control instructions using multi-agent deep reinforcement learning and model predictive control optimization strategies through a cloud decision-making layer, executes instructions and feeds back states through a device execution layer, and simultaneously adopts a ladder optimization strategy, dynamic priority and closed-loop rolling optimization mechanism through the edge computing layer. In this way, energy-saving optimization of the indoor skiing field heating ventilation system is achieved, environmental comfort is ensured, and system operation efficiency, stability and economy are improved. The following will be described and introduced through multiple embodiments with reference to the drawings.

[0024] Figure 1 A system architecture diagram of the energy-saving optimization system of the indoor skiing field heating ventilation air conditioning provided by the embodiment of the present application is shown in FIG. 1. Figure 1 , the energy-saving optimization system of the indoor skiing field heating ventilation air conditioning comprises: A multi-source sensing layer for collecting multi-source data of the indoor skiing field environment; the multi-source data includes weather, passenger flow, temperature and humidity, and device state; An edge computing layer is configured to perform data fusion processing on the multi-source data and generate a load prediction result through a load prediction mechanism. A cloud decision layer is connected with the edge computing layer, configured to receive the load prediction result transmitted by the edge computing layer and the device state fed back by the device execution layer, generate an optimized control instruction based on a multi-agent deep reinforcement learning and a model predictive control optimization strategy. A device execution layer is connected with the cloud decision layer, including HVAC execution devices, configured to receive and execute the optimized control instruction, and feed back the device state to the edge computing layer and the cloud decision layer.

[0025] Specifically, referring to Figure 1 The system adopts a hierarchical architecture design and is composed of four core layers to form a complete closed-loop control system. The first layer is a multi-source perception layer, which is responsible for collecting various data in the indoor ski field environment, including meteorological data, passenger flow data, environmental temperature and humidity data, and the running state data of the HVAC system devices. The second layer is an edge computing layer connected with the multi-source perception layer, and the core function is to fuse the multi-source data transmitted from the upper layer and generate a load prediction result through an embedded load prediction mechanism. The third layer is a cloud decision layer connected with the edge computing layer, which is responsible for receiving the load prediction result from the edge computing layer and the device state feedback data from the device execution layer. This layer adopts a multi-agent deep reinforcement learning (MADRL, Multi-Agent Deep Reinforcement Learning) and model predictive control (MPC, Model Predictive Control) combined optimization strategy to generate the optimal control instruction. The fourth layer is a device execution layer connected with the cloud decision layer, including chiller units, water pumps, cooling towers, and air valves, etc. The function of this layer is to receive and execute the optimized control instruction issued by the cloud decision layer, and at the same time, feed back the real-time running state data of various execution devices to the edge computing layer and the cloud decision layer. The layers are bidirectionally transmitted and interacted through raw data, state feedback and control instructions, and cooperatively realize the energy-saving optimized operation of the HVAC system of the indoor ski field.

[0026] The present application realizes the efficient and energy-saving operation of the HVAC system of the indoor ski field by constructing a multi-layer intelligent optimization architecture. The system comprehensively collects environmental data through the multi-source perception layer, performs real-time data fusion and load prediction through the edge computing layer, generates control instructions through the cloud decision layer adopting the multi-agent deep reinforcement learning and model predictive control collaborative optimization, and accurately executes the instructions and feeds back the running state through the device execution layer. This hierarchical and collaborative closed-loop control mechanism improves the load prediction accuracy of the system, realizes the dynamic collaborative optimization of the devices, effectively reduces the energy loss, and at the same time ensures the environmental comfort of the ski area and the supporting area, improves the system operation efficiency and economy.

[0027] In a preferred embodiment of the present application, the multi-source perception layer includes a weather data acquisition unit, a passenger flow data acquisition unit, a temperature and humidity data acquisition unit, and a device status data acquisition unit. The weather data acquisition unit is used to obtain outdoor temperature, humidity, and solar radiation intensity data. These information is crucial for predicting the cold load of the indoor ski resort and adjusting the air conditioning system to adapt to external weather changes.

[0028] The passenger flow data acquisition unit uses three technical means: gate counting, Wi-Fi probe MAC address identification, and video monitoring image analysis to real-time statistics of the passenger flow in the ski resort, with a sampling frequency of 5 minutes / time to capture the dynamic characteristics of passenger flow changes.

[0029] The temperature and humidity data acquisition unit is deployed in different areas of the ski resort, responsible for continuous monitoring and recording of temperature and humidity data in each area, uploading data every 2 minutes to ensure that the system can respond to changes in indoor environment in a timely manner.

[0030] The device status data acquisition unit is responsible for collecting the operating parameters of key equipment such as refrigeration units, cooling towers, circulating water pumps, and air valves, so that the system can monitor the equipment status in real time, optimize equipment operation, and provide fault warning or diagnosis when necessary.

[0031] Based on data acquisition, the system also accesses multi-dimensional data streams, including historical energy consumption data, high-precision short-term weather forecasts, real-time passenger flow statistics data, and indoor environmental sensor data. The integration of these data streams enables the system to more comprehensively understand the operation status and environmental conditions of the ski resort, providing more accurate and effective support for energy-saving optimization of the air conditioning system.

[0032] In a preferred embodiment of the present application, the edge computing layer includes a load prediction module, which includes a data preprocessing unit, a hybrid deep learning model construction unit, and a dynamic weight self-adaptive unit.

[0033] The data preprocessing unit is used to receive multi-source data collected by the multi-source perception layer, specifically using the Isolation Forest algorithm to detect and eliminate outliers, removing high-frequency noise through moving average filtering, and using linear interpolation method to supplement missing data, to ensure the integrity and reliability of the data input into the subsequent units.

[0034] The mixed deep learning model construction unit takes a long short-term memory network (LSTM) or a Transformer model as a core framework. The input features of the mixed deep learning model construction unit cover climate features, passenger flow features, and historical load features. Among them, the climate features are obtained after processing meteorological data and include normalized data such as temperature and humidity; the passenger flow features are obtained by processing passenger flow data and include current passenger flow and passenger flow change rate in the past 3 hours; and the historical load features are extracted from historical energy consumption data and specifically include a load sequence in the past 24 hours. In the training stage, the unit selects historical energy consumption data in the past two years and corresponding meteorological, passenger flow, temperature, and humidity data to form a training data set, and divides the training data set into a training set and a validation set according to a 7:3 ratio. The model parameters are iteratively optimized by using an Adam optimizer, the root mean square error (RMSE) is used as the loss function for model training, and the model loss value is converged through 500 rounds of iterative training. At the same time, in the training process, the model training focus is flexibly adjusted in combination with the working condition data corresponding to different seasons and different passenger flow intensities, so as to ensure that the model can stably output load prediction results under various working conditions.

[0035] The dynamic weight adaptive unit adopts a lightweight convolutional neural network (CNN), which dynamically adjusts the historical data weight ω1, the real-time meteorological data weight ω2, and the real-time passenger flow data weight ω3 according to real-time disturbance features, and obtains the final load prediction result through weighted fusion of the three. Among them, the real-time disturbance features include a passenger flow anomaly index and a weather change rate. The passenger flow anomaly index is obtained by calculating the deviation rate of the current passenger flow data from the historical average passenger flow, and is determined as abnormal when the deviation rate is >20%; the weather change rate is obtained by calculating the change slope of the outdoor temperature in the past 1 hour. The unit accurately adjusts the proportion of various weights according to the change of real-time disturbance features. Specifically, when the passenger flow anomaly index is >20%, the real-time passenger flow data weight ω3 is automatically increased from the base value 0.3 to 0.5-0.6; when the weather change rate is >2℃ / h, the real-time meteorological data weight ω2 is increased from the base value 0.4 to 0.5-0.55, and the historical data weight ω1 is correspondingly reduced, and ω1+ω2+ω3=1 is satisfied. The final prediction value is the weighted fusion result of the model output value and the dynamic weight, which can accurately predict the cold load and wet load curves in the next 1-24 hours with a time granularity of 15 minutes.

[0036] By fusing meteorological, passenger flow, and equipment state data, and combining the mixed deep learning model and the dynamic weight adaptive mechanism in the edge computing layer, the system can dynamically fuse historical data and real-time disturbance features, accurately predict future load change trends, and effectively solve the prediction lag problem caused by single data and static models in traditional methods, thereby providing an accurate data basis for system optimization and regulation.

[0037] In a preferred embodiment of the present application, the cloud decision layer comprises a collaborative optimization control module, the collaborative optimization control module comprises a multi-agent deep reinforcement learning unit and a model predictive control optimizer; wherein, The multi-agent deep reinforcement learning unit is used to model the chillers, cooling towers, circulating water pumps and terminal air valves as independent agents respectively, the state space of each agent includes its own operating parameters and the environmental parameters of the corresponding area, the action space is the adjustable control parameter range of itself, the reward function integrates the energy consumption cost and the comfort related parameters, and each agent shares the environmental interaction data through an experience replay pool; The model predictive control (MPC) optimizer adopts a linear time-varying (LTV) model, the state variables of the linear time-varying model include the total cooling capacity of the system, the temperatures of each area and the energy consumption of the equipment, the input variables are the load prediction results output by the edge computing layer, the output variables are the target control parameters corresponding to each agent, and the key parameters of the linear time-varying model are obtained through a combination of offline identification and online updating, the initial parameters are identified based on historical operation data in the offline stage, and the model is dynamically updated by receiving the parameter correction values fed back by the multi-agent deep reinforcement learning unit in the online stage.

[0038] Specifically, the collaborative optimization control module realizes device collaborative optimization and global control through deep integration of multi-agent deep reinforcement learning and model predictive control. The specific implementation process includes agent modeling and multi-agent deep reinforcement learning framework building, and the core devices such as chillers, cooling towers, circulating water pumps and terminal air valves are modeled as independent agents. The state space of each agent includes its own operating parameters such as the outlet water temperature and current of the chiller, the frequency and pressure of the water pump, and the environmental parameters of the corresponding area including temperature and humidity. The action space is the adjustable control parameter, such as the load rate adjustment range of the chiller 50% to 100% and the frequency adjustment range of the water pump 30 Hz to 50 Hz. The reward function is designed as the weighted sum of energy consumption cost and comfort penalty, wherein the energy consumption cost is the product of device power and real-time electricity price, and the comfort penalty is the linear penalty term when the actual temperature deviates from the set temperature by more than plus or minus 0.5 degrees Celsius. Figure 2 The MADRL-MPC collaborative mechanism provided in the embodiment of the present application is shown in the schematic diagram Figure 2 The embodiment adopts a multi-agent deep deterministic policy gradient (MADDPG) algorithm, each agent shares the environmental interaction data through an experience replay pool, and learns the collaborative rules between devices through 1000 rounds of simulation running in the training stage.

[0039] The model predictive control optimizer is configured to use a linear time-varying (LTV) model as a system model, the model state variables include total cold capacity of the system, temperature of each region, and energy consumption of the equipment, the input variable is an output value of the load prediction module, and the output variable is a target control parameter of each equipment. Key parameters of the model, such as system response coefficients and equipment efficiency curve coefficients, are obtained by combining off-line identification and on-line updating. In the off-line stage, the initial parameters are identified by using the least square method based on historical operation data, and in the on-line stage, the parameter correction value fed back by the multi-agent deep reinforcement learning is received every 1 hour, and the model is dynamically updated to ensure the consistency of the model and the actual system. The optimization goal of the model predictive control is to minimize the total energy consumption, while considering the comfort constraint requirement that the temperature fluctuation is not more than 0.3 degrees Celsius, and the interior point method is used to quickly solve the optimal control sequence, and the solving period is 5 minutes each time.

[0040] The multi-agent deep reinforcement learning and the model predictive control are combined in the cloud decision layer, the key equipment such as the refrigeration unit, the water pump and the cooling tower is constructed into an intelligent agent with autonomous decision and cooperation capability, dynamic cooperation between the equipment and global energy consumption optimization are realized, and the problems of energy consumption conflict and low overall energy efficiency caused by independent operation of the equipment in the traditional control are solved.

[0041] In a preferred embodiment of the present application, the collaborative optimization control module further comprises a collaborative working unit; the collaborative working unit is used for realizing the collaborative work of the multi-agent deep reinforcement learning unit and the model predictive control optimizer; a load curve output by the load prediction module is used as a feedforward disturbance input of the model predictive control optimizer; the multi-agent deep reinforcement learning unit realizes real-time sensing of the system state, on-line learning of the change of the equipment characteristics through an actor-critic mechanism, and output of a model parameter correction value to the model predictive control optimizer; the model predictive control optimizer solves an optimal control sequence based on an updated linear time-varying (LTV) model, and delivers the control sequence to each intelligent agent as a reference constraint; each intelligent agent combines the learning experience to realize local strategy fine-tuning in the reference constraint range, generates an execution-level control instruction, and simultaneously feeds back actual operation data to an experience replay pool for next round learning and optimization.

[0042] Specifically, the load curve output by the load prediction module is used as the input basis for the model predictive control optimizer to carry out feedforward disturbance calculation. The multi-agent deep reinforcement learning unit continuously and real-timely perceives the system state, and tracks the dynamic changes of the characteristics of the equipment online with the help of the Actor-Critic mechanism, and outputs data for correcting the parameters such as the bias compensation value of the equipment efficiency coefficient to the model predictive control optimizer every 30 minutes. The model predictive control optimizer solves the optimal control sequence applicable to the future 1 hour based on the received and updated linear time-varying (LTV) model, and the optimal control sequence covers the set temperature of the refrigeration unit, the frequency of the water pump and the like, and then the control sequence is issued to each agent as a reference constraint. Each agent combines the accumulated learning experience to implement fine tuning of the local strategy within the range defined by the reference constraint, for example, to perform parameter adjustment with an amplitude of ±5%, and finally generates a control instruction that can be directly executed. At the same time, the agent feeds back the data generated during actual operation to the experience replay pool to provide support for the next round of learning optimization, so as to form a closed-loop iterative collaborative control mechanism to ensure that the system realizes efficient and accurate collaborative regulation under dynamic working conditions.

[0043] The present application realizes efficient and accurate control of the HVAC system through the collaborative optimization mechanism of multi-agent deep reinforcement learning and model predictive control. The scheme uses load prediction data as feedforward input, learns the changes of equipment characteristics in real time through the Actor-Critic mechanism, and dynamically corrects the model parameters every 30 minutes. The model predictive control solves the optimal control sequence based on the updated linear time-varying model, and each agent performs ±5% parameter fine tuning within the constraint range, which not only ensures the global optimization effect, but also fully gives play to the adaptive ability of the equipment. The actual operation data is fed back to the experience replay pool to form a closed-loop learning mechanism, which significantly improves the control accuracy and energy efficiency of the system under dynamic working conditions, and realizes the dual goals of energy saving and comfort guarantee.

[0044] In a preferred embodiment of the present application, the edge computing layer further comprises a ladder optimization execution module, and the ladder optimization execution module comprises a five-stage optimization strategy; the five-stage optimization strategy comprises: A single equipment independent optimization stage, based on the energy efficiency curve characteristics of the refrigeration unit, combined with real-time load demand, the optimal load rate is found through quadratic function fitting, usually 70% to 85%, and the unit output power is adjusted accordingly; at the same time, according to the similarity law, the circulating water pump frequency is adjusted according to the actual flow demand to ensure that the water pump operates in the high efficiency zone with an efficiency not less than 75%; In the small system collaborative optimization stage, the collaborative relationship between the refrigeration host, the water pump and the cooling tower is optimized through the flow and temperature difference matching algorithm. Under the premise of ensuring the outlet water temperature of the host, the fan frequency of the cooling tower is dynamically adjusted to control the inlet water temperature of the cooling water, and the pump frequency is adjusted to control the flow, so that the total energy consumption of the small system is reduced by 5% to 8%; In the regional end optimization stage, each temperature zone in the ski field is taken as a unit, including the primary road, the intermediate road and the rest area, etc. Based on the real-time temperature and humidity sensor data, a proportional-integral-derivative control algorithm is used to adjust the opening degree of the end air supply valve, the adjustment range is 0 to 100%, and the reheater power is adjusted to ensure the uniformity of the temperature and humidity in the region, the temperature deviation is not more than ±0.5 degrees Celsius, and the humidity deviation is not more than ±3%. In the global optimization stage of the whole system, all subsystem operation data are integrated, and based on the global optimal target output by the multi-agent deep reinforcement learning and model predictive control optimization module, the linear programming method is used to balance the load distribution of each region, avoid overloading or inefficient operation of the equipment, and realize the minimization of the total energy consumption of the system. In the economic fine-tuning decision stage, real-time electricity price signals are introduced, including peak electricity price, flat electricity price and valley electricity price, and a device life loss model is combined, which calculates the fatigue loss based on the device running time and start-stop number. On the basis of global optimization, the control strategy is fine-tuned through cost-benefit analysis, for example, the temperature control accuracy requirement of non-core areas is appropriately relaxed during the peak electricity price period, allowing a deviation of ±0.8 degrees Celsius, and the comfort of the skiing area is preferentially guaranteed, so as to reduce the system operation cost.

[0045] The present application realizes the fine energy efficiency management of the heating, ventilation and air conditioning system through the five-stage ladder optimization strategy. The single device optimization ensures that the refrigeration unit and the water pump operate in the high efficiency interval; the small system collaboration reduces the energy consumption by 5% to 8% through flow and temperature difference matching; the regional end optimization ensures the uniformity of the temperature and humidity in each temperature zone; the global optimization realizes the minimization of the total energy consumption of the system based on MADRL-MPC; and the economic fine-tuning combines the real-time electricity price and the device life model to intelligently adjust the temperature control strategy during the peak electricity price period. This progressive optimization method effectively improves the system energy efficiency and reduces the operation cost under the premise of ensuring the comfort of the skiing area.

[0046] In a preferred embodiment of the present application, the ladder optimization execution module further includes a dynamic priority mechanism execution unit; the dynamic priority mechanism execution unit is used for dynamically adjusting the optimization strategy according to the real-time monitored passenger flow state and time period characteristics: When the passenger flow density of the skiing area is greater than a set threshold, a high load response mode is triggered, the global optimization of the whole system and the economic fine-tuning decision stage are preferentially executed, and the control period is shortened; during the night low load period, the energy saving mode is switched to, the single device independent optimization and the small system collaborative optimization stage are focused on, and the control period is prolonged.

[0047] In a specific implementation process, the system monitors the passenger flow state and time period characteristics in real time. When the passenger flow density of the ski area exceeds 2 people per square meter, which is determined by video analysis technology, the system triggers the high load response mode, preferentially executes the fourth stage of global optimization of the whole system and the fifth stage of economic fine-tuning decision, and shortens the control cycle to 2 minutes each time to realize rapid response to load changes. During the low load period from 22:00 to 06:00 the next day, the system automatically switches to the energy-saving mode, focuses on the first stage of single-device independent optimization and the second stage of small-system collaborative optimization, and extends the control cycle to 10 minutes each time to ensure efficient operation of the basic working condition.

[0048] The five-stage optimization strategy and dynamic priority mechanism realized by the ladder optimization execution module enable the system to adaptively adjust the optimization focus and control cycle according to real-time passenger flow density, time period characteristics and other working condition changes, realize hierarchical fine-tuning control from the device level to the system level, and improve the response speed, stability and economy of the system under different operating scenarios.

[0049] In a preferred embodiment of the present application, the ladder optimization execution module further comprises a closed-loop rolling optimization unit; the closed-loop rolling optimization unit is used to: issue the optimized control instructions to the field actuators through the Internet of Things system, and receive the actual action state feedback from the actuators; periodically compare the deviation between the actual running data and the optimization target, and trigger the emergency correction mechanism when the deviation exceeds the set threshold, re-execute the optimization process of the corresponding stage, and realize closed-loop control.

[0050] In a specific implementation process, the optimized control instructions are issued to the field actuators including valves and frequency converters through the Internet of Things system, and the actuators feed back the actual action state to the system. The system compares the deviation between the actual running data and the optimization target every 5 minutes, and if the deviation is greater than 10%, the emergency correction mechanism is triggered, the optimization process of the corresponding stage is re-executed, and a complete closed-loop control system is formed. This mechanism ensures that the system can respond to changes in operating state in a timely manner, maintain optimization effect and improve system stability.

[0051] The present application realizes efficient and energy-saving operation of the heating, ventilation and air conditioning system of the indoor ski field by constructing a multi-layer intelligent optimization architecture. The system comprehensively collects environmental data through the multi-source perception layer, performs real-time data fusion and accurate load prediction through the edge computing layer, generates control instructions through the cloud decision layer by adopting multi-agent deep reinforcement learning and model predictive control collaborative optimization, and accurately executes instructions and feeds back the operating state through the device execution layer. A five-stage ladder optimization strategy is adopted to realize fine regulation and control from the device level to the system level, and a dynamic priority mechanism is combined to adaptively adjust the optimization strategy. Through closed-loop rolling optimization, the system can ensure real-time response to working condition changes, significantly improve load prediction accuracy and device collaboration efficiency, effectively reduce system energy consumption by 5%-8% under the premise of ensuring the comfort of the skiing area, and improve operation stability and economy.

[0052] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An energy-saving optimization system for indoor ski resort HVAC, characterized in that, include: The multi-source sensing layer is used to collect multi-source data on the indoor ski resort environment; the multi-source data includes weather, passenger flow, temperature and humidity, and equipment status. The edge computing layer is used to perform data fusion processing on the multi-source data and generate load forecasting results through a load forecasting mechanism; The cloud-based decision layer connects to the edge computing layer and receives load prediction results transmitted from the edge computing layer and device status feedback from the device execution layer. Based on multi-agent deep reinforcement learning and model predictive control optimization strategies, it generates optimized control commands. The device execution layer, connected to the cloud decision layer, includes HVAC execution devices for receiving and executing the optimized control commands, and simultaneously feeding back the device status to the edge computing layer and the cloud decision layer.

2. The indoor ski resort HVAC energy-saving optimization system according to claim 1, characterized in that, The multi-source sensing layer includes a meteorological data acquisition unit, a passenger flow data acquisition unit, a temperature and humidity data acquisition unit, and an equipment status data acquisition unit. The meteorological data acquisition unit is used to collect outdoor temperature, humidity and solar radiation intensity data; The passenger flow data acquisition unit is used to collect passenger flow data through gate counting, Wi-Fi probe MAC address identification, and video surveillance image analysis; The temperature and humidity data acquisition units are distributed in different areas of the indoor ski resort and are used to collect temperature and humidity data for each area. The equipment status data acquisition unit is used to collect operating parameters of the refrigeration unit, cooling tower, circulating water pump, and air valve.

3. The indoor ski resort HVAC energy-saving optimization system according to claim 1, characterized in that, The edge computing layer includes a load prediction module, which comprises a data preprocessing unit, a hybrid deep learning model construction unit, and a dynamic weight adaptation unit; wherein... The data preprocessing unit is used to receive multi-source data collected by the multi-source sensing layer, use the isolated forest algorithm to detect and remove outliers, remove high-frequency noise by moving average filtering, and then use linear interpolation to supplement missing data. The hybrid deep learning model building unit uses a long short-term memory network or a Transformer model as its core framework. The input features include climate features, passenger flow features, and historical load features. The climate features are obtained by processing meteorological data, the passenger flow features are obtained by processing passenger flow data, and the historical load features are obtained by extracting historical energy consumption data. The dynamic weight adaptive unit employs a lightweight convolutional neural network to dynamically adjust the weights of historical data, real-time meteorological data, and real-time passenger flow data based on real-time disturbance characteristics, and obtains the load prediction result through weighted fusion of the three.

4. The indoor ski resort HVAC energy-saving optimization system according to claim 3, characterized in that, The hybrid deep learning model building unit selects historical energy consumption data and corresponding meteorological, passenger flow, temperature and humidity data to form a training dataset during the training phase, and divides the training dataset into a training set and a validation set. The hybrid deep learning model building unit uses the Adam optimizer to iteratively optimize the model parameters, uses the root mean square error as the loss function for model training, and converges the model loss value through multiple iterations. During the training process, the hybrid deep learning model building unit adjusts the training focus of the model by combining operating data corresponding to different seasons and different passenger flow intensities, so as to ensure that the model can stably output load prediction results under various operating conditions.

5. The indoor ski resort HVAC energy-saving optimization system according to claim 3, characterized in that, The real-time disturbance characteristics include passenger flow anomaly index and weather change rate; The passenger flow anomaly index is obtained by calculating the deviation rate between the current passenger flow data and the historical average passenger flow for the same period. The rate of weather change is obtained by calculating the slope of outdoor temperature change within a set time period. The dynamic weight adaptive unit adjusts the proportions of historical data weight, real-time meteorological data weight, and real-time passenger flow data weight according to changes in real-time disturbance characteristics, so that the weighted fusion load forecast result is adapted to the current operating conditions.

6. The indoor ski resort HVAC energy-saving optimization system according to claim 1, characterized in that, The cloud-based decision-making layer includes a collaborative optimization control module, which comprises a multi-agent deep reinforcement learning unit and a model prediction control optimizer; wherein... The multi-agent deep reinforcement learning unit is used to model the terminal air valves of the cooling tower circulating water pump of the chiller unit as independent agents. The state space of each agent includes its own operating parameters and the environmental parameters of the corresponding area. The action space is the range of its own adjustable control parameters. The reward function integrates energy consumption cost and comfort-related parameters. Each agent shares environmental interaction data through the experience replay pool. The model predictive control optimizer adopts a linear time-varying LTV model. The state variables of the linear time-varying LTV model include the total cooling capacity of the system, the temperature of each region, and the energy consumption of the equipment. The input variables are the load prediction results output by the edge computing layer, and the output variables are the target control parameters corresponding to each agent. The key parameters of the linear time-varying model are obtained by combining offline identification and online updating. In the offline stage, the initial parameters are identified based on historical operating data. In the online stage, the model is dynamically updated by receiving parameter correction values ​​fed back by the multi-agent deep reinforcement learning unit.

7. The indoor ski resort HVAC energy-saving optimization system according to claim 6, characterized in that, The collaborative optimization control module also includes a collaborative working unit; the collaborative working unit is used to realize the collaborative work between the multi-agent deep reinforcement learning unit and the model predictive control optimizer. The load curve output by the load forecasting module serves as the feedforward perturbation input for the model predictive control optimizer; the multi-agent deep reinforcement learning unit perceives the system state in real time, learns the changes in equipment characteristics online through the actor-critic mechanism, and outputs model parameter correction values ​​to the model predictive control optimizer. The model predictive control optimizer solves for the optimal control sequence based on the updated linear time-varying LTV model and distributes the control sequence as a reference constraint to each agent. Each agent, based on its own learning experience, fine-tunes its local strategy within the reference constraints, generates execution-level control instructions, and feeds back the actual running data to the experience replay pool for the next round of learning and optimization.

8. The indoor ski resort HVAC energy-saving optimization system according to claim 1, characterized in that, The edge computing layer also includes a ladder optimization execution module, which includes a five-stage optimization strategy; The five-stage optimization strategy includes: In the single-device independent optimization stage, the optimal load rate is found and the output power is adjusted based on the energy efficiency curve of the chiller unit and the real-time load demand. Based on the similarity law, the frequency of the circulating water pump is adjusted according to the actual flow demand to make it operate in the high-efficiency zone. In the small system collaborative optimization stage, the collaborative relationship of the chiller, water pump and cooling tower system is optimized by the flow-temperature difference matching algorithm, and the frequency of cooling tower fan and water pump is dynamically adjusted to reduce the total energy consumption of the small system. In the regional-level end-of-line optimization stage, based on real-time temperature and humidity sensor data, the opening degree of the end-of-line air outlet valve and the power of the reheater are adjusted through proportional-integral-derivative PID control algorithm for each temperature zone of the ski resort to ensure the uniformity of temperature and humidity within the area. In the global optimization phase of the entire system, data from all subsystems are integrated, and the global optimal objective output by the multi-agent deep reinforcement learning and model predictive control optimization module is used to minimize the total energy consumption of the system by balancing the load distribution of each region through linear programming. In the economic fine-tuning decision-making stage, real-time electricity price signals and equipment life loss models are introduced, and the control strategy is fine-tuned through cost-benefit analysis based on global optimization.

9. The indoor ski resort HVAC energy-saving optimization system according to claim 8, characterized in that, The tiered optimization execution module also includes a dynamic priority mechanism execution unit; the dynamic priority mechanism execution unit is used to dynamically adjust the optimization strategy based on the real-time monitored passenger flow status and time period characteristics. When the passenger flow density in the ski area exceeds the set threshold, a high-load response mode is triggered, prioritizing the global optimization and economic fine-tuning decision-making stages of the entire system to shorten the control cycle. During low-load periods at night, switch to energy-saving mode, focusing on the independent optimization of single devices and the collaborative optimization of small systems, and extend the control cycle.

10. The indoor ski resort HVAC energy-saving optimization system according to claim 8, characterized in that, The ladder optimization execution module further includes a closed-loop rolling optimization unit; the closed-loop rolling optimization unit is used for: The optimized control commands are sent to the field actuators through the Internet of Things system, and the actual action status is received from the actuators. Regularly compare the deviation between actual operating data and optimization targets. When the deviation exceeds a set threshold, trigger an emergency correction mechanism to re-execute the corresponding stage of the optimization process, thereby achieving closed-loop control.

Images